EM signal integrity via neural network analysis for the RFX-mod experiment

Abstract

The RFX-mod electromagnetic measurement system is constituted of 744 independent probes whose signals are electronically conditioned by an integration/amplification section. During experimental sessions the probes integrity is controlled by a series of post-shot softwares which determine if a probe is still working or not and correct off-sets and drifts, but no method, apart from the visual inspection of a signal, is available to recognize if the corresponding channel in the integration/amplification section is about to break. In order to overcome this lack a neural network approach has been applied. The neural network implemented here is built performing a geometrical synthesis of a supervised Multi Layer Perceptron, then the trained net is used to predict a possible failure of the corresponding channel in the integration/amplification section. To perform the prediction the neural network is used as a non linear regressor, the synaptic weights of the trained net can be considered as a neural transform of the system, the variation of those weights in the test phase is symptom that the channel is not working properly. The procedure has been tested on a subset of electromagnetic signals and in this paper the results are presented.